RESUMEN
Alzheimer's disease (AD) is a prevalent type of dementia that can cause neurological brain disorders, poor decision making, impaired memory, mood swings, unstable emotions, and personality change. Deep neural networks are proficient in classifying Alzheimer's disease based on MRI images. This classification assists human experts in diagnosing AD and predicts its future progression. The paper proposes various Deep Neural Networks (DNN) for early AD detection to save cost and time for doctors, radiologists, and caregivers. A 3330-image-based Kaggle dataset is used to train the DNN, including 52 images of AD, 717 images of Mild Cognitive Impairment (MCI), and the remaining images of Cognitive Normal (CN). Stratified partitioning splits the dataset into 80% and 20% proportions for training and validation datasets. Proposed models include DenseNet169, DenseNet201, and Res- Net152 DNNs with additional three fully-connected layers and softmax and Kullback Leibler Divergence (KLD) loss function. These models are trained considering pre-trained, partially pre-trained, and fully re-trained extended base models. The KLD loss function reduces the error and increases accuracy for all models. The partially pre-trained DenseNet201 model outperformed all the other models. DenseNet201 gives the highest accuracy of 99.98% for training, 99.07% for validation, and 95.66% for test datasets. The DenseNet201 model has the highest accuracy in comparison to other state-of-artmethods.
